CO

MH

a

ARR1AA

JDQ

KNTPR

1

tantacoc

CC2s

0

Journal of Economic Behavior & Organization 118 (2015) 199–214

Contents lists available at ScienceDirect

Journal of Economic Behavior & Organization

j ourna l h om epa ge: w ww.elsev ier .com/ locate / jebo

atastrophes and time preference: Evidence from the Indiancean Earthquake�

ichael Callen ∗

arvard Kennedy School, United States

r t i c l e i n f o

rticle history:eceived 27 February 2014eceived in revised form7 November 2014ccepted 22 February 2015vailable online 18 March 2015

EL classification:9154

eywords:atural disastersime preferencereference stabilityegression discontinuity

a b s t r a c t

We provide evidence suggesting that exposure to the Indian Ocean Earthquake tsunamiincreased patience in a sample of Sri Lankan wage workers. We develop a framework tocharacterize the various channels through which disaster exposure could affect measuresof patience. Drawing on this framework, we show that a battery of empirical tests supportthe argument that the increase in measured patience reflects a change in time preferenceand not selective exposure to the event, migration related to the tsunami, or other changesin the economic environment which affect experimental patience measures. The resultshave implications for policies aimed at disaster recovery and for the literature linking lifeevents to economic preferences.

© 2015 Elsevier B.V. All rights reserved.

. Introduction

Time preference features centrally in theories of consumer optimization, economic growth, and interest rate determina-ion. Intertemporal preferences, additionally, are of key relevance for development economics. Capital investments, healthnd education investments made for children, time spent learning, and other accumulation decisions directly affect eco-omic success. Evaluating whether time preference responds to disasters may therefore be important in understandinghe economic legacies of these events. Additionally, if major life events impact preferences, then understanding this link-ge is relevant to efforts to account for the sources of heterogeneity in economic taste. In this paper, we provide evidence

onsistent with measured discount factors increasing in response to the experience of a major natural disaster. Measuresf time preference, in addition to measuring the rate of time preference, also reflect beliefs about the future, backgroundonsumption, and intertemporal arbitrage opportunities. Using a simple model that relates underlying preferences and this

� I am grateful to Suresh de Mel, David McKenzie, and Christopher Woodruff for kindly sharing their data and to James Andreoni, Eli Berman, Alessandraassar, Marjorie Flavin, Gordon Dahl, Peter Kuhn, Leah Nelson, Sarah Reber, Charles Sprenger, Craig McIntosh, Christopher Woodruff, Hal White, andarrie Wenjing Xu, as well as seminar participants at the University of San Francisco, the 2011 Pacific Development Conference at UC Berkeley, and the011 Southern California Conference in Applied Microeconomics at Claremont McKenna for many helpful insights. Callen acknowledges graduate studentupport from the Air Force Office of Scientific Research (AFOSR) under award FA9550-09-1-0314. All mistakes are my own.∗ Correspondence to: Harvard University, Harvard Kennedy School, 79 JFK Street, Cambridge, MA 02138, United States. Tel.: +1 617 495 9965.

E-mail address: michael callen@hks.harvard.edu

http://dx.doi.org/10.1016/j.jebo.2015.02.019167-2681/© 2015 Elsevier B.V. All rights reserved.

200 M. Callen / Journal of Economic Behavior & Organization 118 (2015) 199–214

set of confounds to revealed laboratory measures, we show that the increase in patience we observe is only consistent witha change in underlying preferences or a large, and implausible, decrease in the return to capital from the tsunami.

In the set of economic motivations for discounting identified in early research (Rae, 1834; Jevons, 1888, 1905; Senior,1836; Böhm-Bawerk, 1889), the specific taste for trading off utility over time is sharply distinguished by modern theoryfrom considerations which affect expectations and the intertemporal budget constraint.1 Olson and Bailey (1981), providea clear articulation of this broadly accepted view, which we follow, terming the complete set of reasons for impatience timediscounting and the specific taste for intertemporal utility tradeoffs time preference.

Stigler and Becker (1977) emphasize the importance of this distinction as it allows the “economist [to continue] tosearch for differences in prices or income to explain any differences or changes in behavior,” without having to account fordifferences in taste. More recently, economists have begun to investigate whether tastes respond systematically to economicshocks (Voors et al., 2012; Bchir and Willinger, 2013; Willinger et al., 2013; Cameron and Shah, 2013; Cassar et al., 2013;Malmendier and Nagel, 2011; Meier and Sprenger, forthcoming; Callen et al., 2014) and also if preferences develop as aresult of utility-maximizing investments in preference formation (Becker and Mulligan, 1997).

Our data, collected in regions affected by the Indian Ocean Earthquake (formally called the Sumatra-Andaman earthquake)in Sri Lankan are well-suited to testing whether economic preference responds to catastrophe. The event, which took placeon December 26, 2004 off the west coast of Sumatra, Indonesia devastated household assets. In our sample, 153 of 456wage workers (33.55%) report at least some damage and the median affected individual reported suffering approximately200,000 rupees worth of damage, which is approximately 23.7 times the median monthly wage. The tsunamis triggeredby the earthquake significantly damaged the coasts bordering the Indian Ocean. Waves up to 30 m were recorded and theUN reports 229,866 total individuals lost or missing. In Sri Lanka, 35,322 individuals were killed and 516,150 people weredisplaced.2 The economic and psychological damage wrought by the tsunami was unexpected and severe. It likely affectedhorizons, consumption streams, and made salient the possibility of extreme unanticipated losses. Using a standard measureof willingness to delay consumption, we find that workers affected by the tsunami are more patient than unaffected workers.

More specifically, our estimates are consistent with the tsunami creating an increase in the average monthly discountfactor from about 0.8 to about 0.85 or to about 0.9 depending on the specification. This represents a 1/3rd to 2/3rd of astandard deviation increase in our patience measure. These estimates are robust to a broad set of specifications and arevery similar to regression discontinuity results. Additionally, the effect appears to be enduring. Our data reflect preferenceselicited two and a half years after the Indian Ocean Earthquake. We also find that this effect is largest for individuals who:have no secondary or higher education, are shorter and so are less likely to have received sufficient caloric nourishment aschildren, perform worse on a cognitive test and, who are in the left tail of the patience distribution.3 While this only letsus begin to speculate as to the reasons that patience appears to have increased as a result of the tsunami, these results areconsistent with tsunami exposure being a substitute for other inputs which are argued to develop patience (Becker andMulligan, 1997). The direction of the effect, when contrasted against the predictions of our simple theory, and the richnessof our data allow us to argue that our results are not driven by liquidity constraints and other considerations that affectelicited discount rates but that are not part of standard theoretical concept of the rate of time preference.

There are two principal issues confronting attempts to establish the causal effect of disasters on economic preferencesusing data collected after a disaster. The first is that individuals may locate according to their preferences, which we callselective exposure. In our case, patient workers may select areas vulnerable to tsunami inundation. Second, in the wake ofthe disaster, individuals might selectively migrate out of affected zones based on their preferences. We call this selectivemigration.

We lack both longitudinal data tracking individuals before and after the disaster and random assignment and so cannotrule out these confounds. However, several results indicate that these concerns do not account for the increase in patiencethat we observe. First, the implied magnitude of the increase is robust to using variation only within small arbitrary spatialdivisions (fishnet grids) and to a regression discontinuity design based on distance from the high water mark, which shouldcontrol of additional differences that relate to proximity to the disaster. Second, we find affected and unaffected workers arebalanced on a broad range of time-invariant observable characteristics. Third, the share of respondents living in the samecommunity since birth is not related to tsunami exposure, indicating no major effect on migration patterns in our sampletwo years after the event. Fourth, looking at the part of our sample which is vulnerable to inundation but escaped exposurebecause of the wave’s direction, we find no evidence that more patient individuals are disproportionately vulnerable (con-trary to selective exposure) and find comparable shares of individuals living in the same community since birth (contrary toselective migration).

To establish that the change we observe reflects a change in preferences we also need to consider a set of competingexplanations. To precisely specify potential confounds, we develop a framework to characterize the various channels through

which disaster exposure could affect measures of patience. It is possible that because survey questions which ask aboutconsumption trade-offs over time can only measure time discounting, which reflects a broader range of considerations thanthe parameter we are interested – time preference, the effect we estimate does not truly reflect a change in tastes. To counter

1 Frederick et al. (2002) provide an authoritative review of the intellectual history of intertemporal choice.2 See http://en.wikipedia.org/wiki/2004 Indian Ocean earthquake.3 The full set of these results are available in Appendix Section A2.

tctaidtprti

dmpwSc

2

2

te2fcFpltdmp

2

fwtermuw(lpfic

ha

c

M. Callen / Journal of Economic Behavior & Organization 118 (2015) 199–214 201

his concern, we model the survey respondents’ problem and show that an increase in patience is consistent only with ahange in time preference. Second, it may be that the revealed willingness to delay consumption merely reflects a demando replace assets demolished during the disaster. We find that the effect of the disaster on preferences does not differ byge, childhood poverty status, debt levels, or wages and we do not find any effect of the percentage of assets replace or thentensity of the damage on preferences after controlling for whether an individual was affected at all. Moreover, a 1-monthelay to a payment may be negligible in terms of life cycle saving two and one-half years after the event. Last, it may behat the large-scale aid response to the tsunami may have either relaxed credit constraints or may have flattened the timeath of consumption for beneficiaries by making them richer. However, we find a similar increase for individuals even afterestricting our sample to individuals who received less than a day’s wage worth of aid support. In light of the result fromhe model that an increase in measured patience is only consistent with an actual change in time preference, we interpret thencrease in the measure due to the disaster as reflecting an increase in patience.

The remainder of the paper is organized as follows. Section 2 reviews the distinction, discussed above, between timeiscounting and time preference and also relates our study to the empirical literature that relates intertemporal decision-aking to economic success. Section 3 builds on the potential links identified by the literature and derives theoretical

redictions about how changes in survival expectations and differential marginal utility from changing consumption levelsill affect responses to the survey question we use to elicit time preference. Section 4 provides an overview of our data.

ection 5 reviews our empirical strategy. Section 6 documents the effect of the catastrophe on preferences, and Section 7oncludes.

. Literature

.1. Time discounting and time preference

Olson and Bailey (1981) and Frederick et al. (2002) draw a sharp theoretical distinction between time discounting andime preference. This distinction is key to our analysis because a well-known criticism of experimental time preferencelicitation techniques is that they confound several factors which affect intertemporal decisions (Andreoni and Sprenger,012). Specifically, standard preference elicitation techniques measure time discounting or the complete set of reasonsor discounting the future, including uncertainty, changing tastes, and differential marginal utility arising from changingonsumption levels. By contrast, time preference refers exclusively to the preference for immediate utility over delayed utility.rederick et al. (2002) review the set of considerations which cause elicited preferences to diverge from the rate of timereference. These are (a) intertemporal arbitrage; (b) diminishing marginal utility; (c) uncertainty about the future; (d) price-

evel inflation; (e) expectations of changing utility; and (f) considerations of habit formation. Harrison et al. (2008) confirmhe empirical relevance of diminishing marginal utility and show that it generates a large downward bias in measurediscount factors.4 For the purpose of clarity, we use the terms discounting, patience, impatience, discount factors, andeasured discount factors to refer to time discounting and we reserve time preference specifically for the pure rate of time

reference.

.2. Life experience, preferences, and economic success

Our findings relate to two additional lines of research. The first is the empirical literature linking life experience toundamental and enduring changes in preferences and behavior.5 Malmendier and Nagel (2011) show that individualsho have lived through periods of exceptional stock market performance exhibit less risk aversion and that birth cohorts

hat experienced the Great Depression are much less likely to participate in the stock market. Callen et al. (2014) providevidence that exposure to violence in Afghanistan, while it does not lastingly alter preferences per se, does appear to makeisk preferences more vulnerable to the recollection of prior violent episodes. Meier and Sprenger (forthcoming) show thateasured time preferences do not change in response to more common shocks, such as the loss of employment, among

sers of a volunteer tax assistance site in Boston. Blattman and Annan (2010) show that childhood conscripts in Ugandaho experience the most severe violence are more likely to report psychological distress as adults, and Alesina and Giuliano

2010) provide evidence that historical experiences and personal histories shape preferences for social equity. Along theseines, Fisman et al. (2013) find that the “Great Recession” increased both laboratory measures of selfishness and of the relativereference for equity over equality. Castillo and Carter (2011) look at the effects of Hurricane Mitch in rural Honduras. Theynd that intermediate shocks increase coordination in an anonymous interaction game, while extreme events reduce suchoordination.

The second line of research focuses on the economic implications of disasters and of responses to disasters. Economistsave investigated, for example, how fertility responds to disasters (Finlay, 2009; Pörtner, 2008), whether aid in response to

disaster improves perceptions of the aid provider (Andrabi and Das, 2010) and how to best target reconstruction aid to

4 Andreoni and Sprenger (2012) provide an alternative elicitation procedure which can directly measure time preference.5 We point interested readers to Chuang and Schechter (2014), who provide an excellent and comprehensive review of this literature. Here, we review

ontributions most directly linked to our study.

202 M. Callen / Journal of Economic Behavior & Organization 118 (2015) 199–214

rebuild microenterprises (de Mel et al., 2011). The papers most closely linked to this study are Cameron and Shah (2013),Cassar et al. (2013), and Aubert and Aubert (2013). These three papers look for changes in experimentally elicited preferencemeasures as a result of experiencing disasters. Cameron and Shah (2013) and Cassar et al. (2013) find an increase in riskaversion using measures of risk aversion elicited through incentivized choice experiments. Aubert and Aubert (2013) find adecrease in risk aversion in the losses domain. Andrabi and Das (2010), in an ancillary result, find a similar change as a resultof the 2005 earthquake in Northern Pakistan for measurements that are not incentivized. While we have a similar measureof risk-aversion elicited through incentivized Holt–Laury Multiple Price Lists, our focus, for reasons we describe below, is ontime preference. Given the attention paid by both Cameron and Shah (2013) and Cassar et al. (2013) to issues of respondentcomprehension and to careful measurement, we direct readers interested in the risk preference effects of disasters to thosestudies.6

The primary difference between the current study and both Cameron and Shah (2013) and Cassar et al. (2013) is that ourfocus is on time preference rather than risk preference. Our focus on time preference, as we discuss in detail in Section 3 below,permits our focus on separately identifying changes in time preference from other changes to the economic environment,such as changes to the time path of consumption and to subjective survival expectations (those two papers acknowledgethe relevance of this distinction).

As in Cameron and Shah (2013) and Cassar et al. (2013), we establish a change in preference measures using a t-test ofmeans between affected and unaffected survey respondents. A secondary difference, however, is that our data on preciserespondent locations allow us to additionally identify the effect using regression discontinuity and also allow us to directlytest the identifying assumptions common to our paper and to Cameron and Shah (2013) and Cassar et al. (2013) – whetherselective migration and selective exposure are present in our sample. Based on this, we argue for a causal interpretationof the difference in means. Our information on respondent locations also allows us to establish that our result is robustto within variation for small arbitrary fishnet grids. We interpret the robustness of the core result, the similarity of theresult estimated through regression discontinuity, and our failure to reject selection of vulnerability based on preferencesas strongly supporting the empirical strategy used in all three papers. Bchir and Willinger (2013) and Willinger et al. (2013)provide results on the effects of lahars – mudflows that emanate from volcanoes – on risk and time preferences. Bchir andWillinger (2013) examine the effects of lahars in Peru and Willinger et al. (2013) examine the impacts in Java. The approach,in both papers, overcomes the selective migration issue, as they observe preferences both before and after the lahars. Bchirand Willinger (2013) find an increase in risk tolerance and impatience. Willinger et al. (2013) find an overall increase in risktolerance and both increases and decreases in time preferences within subjects across time.

In addition to changes in trust, trust worthiness, and MPL measures of risk preference, Cassar et al. (2013) find modestdecreases in patience as a result of the Indian Ocean Earthquake in Thailand for individuals who experience injuries or adeath in the family after controlling for risk preferences.7 The measure which is significally correlated with time preferencein this paper is personal injury or family death, which is distinct from our measure, which is any exposure at all. As weshow in Section 3, this result is consistent both with a change to respondents’ subjective survival probabilities and with anincrease in the marginal utility of consumption due to a decrease in consumption levels resulting from losing a family sourceof income. We therefore view our results as consistent with those of Cassar et al. (2013) as they support the same model.Their measure of damage – injuries or a death in the family – should, according to our theory, generate a different result.

3. Theory

To assess whether the increase in our survey measure of patience reflects a change in time preference and not a responseto the set of extra-time preference considerations listed above, we develop theoretical predictions regarding how theseconsiderations can be expected to change our measure as a result of the tsunami. This distinction is critical given that oursurvey question does not measure time preference directly.

3.1. Predictions of decreased patience

3.1.1. Concave utility and uncertaintyIn our sample, the median affected individual reports suffering approximately 200,000 rupees worth of damage, which is

approximately 23.7 times the median monthly wage. We expect the devastation and the rapid subsequent aid response to put

6 In results available on request, we cannot reject no change in risk preference from the disaster using the same identification strategies we use for ourresult on time preference. This result is consistent to the lack of a difference found in Hurricane Katrina victims one year after the event in Eckel et al.(2009). A potential explanation for our lack of a result is that both Andrabi and Das (2010) and Cassar et al. (2013) find increases in trust as a result ofexperiencing disaster and that Ben-Ner and Putterman (2001) and Karlan (2005) find that risk aversion and trust are negatively correlated so that the trusteffects of the disaster are offsetting the risk preference effects. However, given the broad range of extra-preference confounds that may enter into pricelist elicitations of risk preference, we can only speculate as to the cause of the difference.

7 Given the evidence for random assignment to calamity in this paper as well as in Andrabi and Das (2010), Cameron and Shah (2013), Cassar et al.(2013), it seems likely that the effect on time preference is only significant in specifications which include risk preference measures as a covariate becauseof the importance of risk preference in soaking up noise in measures of time preference and not because risk preference is needed to satisfy conditionalindependence.

itr

YiSpmmtu

r

3

pdtbAftmc

w

btcwtti

P

tt

ctate

M. Callen / Journal of Economic Behavior & Organization 118 (2015) 199–214 203

ndividuals on a steeper (though dramatically lower) consumption path and thereby create incentives to move consumptiono the present. Below, we show how this intuition plays out by considering the optimization problem faced by a surveyespondent.

Additionally, we expect that surviving the tsunami had some effect on individuals subjective beliefs about future payoffs.aari (1965) and Blanchard (1985) provide finite-horizon models that link expected survival probabilities to time discount-

ng. Survival expectations are modeled as a per period probability of dying that is added directly to the rate of time preference.imilarly, Jayachandran and Lleras-Muney (2009) provide a model where the stream of future consumption utility has someositive probability of not realizing because of death during childbirth which depends, in turn, on the prevailing rate ofaternal mortality. Jayachandran and Lleras-Muney (2009) show additionally that a large exogenous reduction in maternalortality in Sri Lanka increased female education and literacy, principally through parental investments in children. Given

his evidence, it is clear that we should take seriously the implications of changing uncertainty and differential marginaltility for the discount rate we impute.

Another parameter that might be affected by the tsunami are interest rates. We extend our theory to consider interestates in Appendix Section A4.

.1.2. Concave utility and uncertainty – predictions from theoryAs we describe more fully in Appendix A6 below, we measure discounting using responses to questions of the form:

“Suppose someone was going to pay you Rupees m six months from now. He/she offers to pay you a lower amount x infive months time. What amount in five months would make you just as happy as receiving Rupees m in six months?”

To see how concave utility and uncertainty affect our imputed measure, consider the respondent’s problem. Because ourredictions are general with respect to the functional form of the discount factor over time, we consider only the exponentialiscounting case. Let u(c) be a concave utility function defined over consumption c and, without loss of generality, u(0) = 0 sohat the respondent receives no utility if they do not consume. For simplicity, we assume a one-month interest rate of zero,ut show below that an increase in the marginal return to capital from asset devastation should decrease discount factors.dditionally, let c0 and c1 be equilibrium consumption today and in one month, m be a fixed reward to be provided in the

uture, p be the probability that an individual is alive one month from today, ı be the discount factor, and x be compensationhat is required to forego some reward in the future. We measure the individual discount factor in our data ı̂ as x/m. Because

is fixed, it is sufficient to consider the predictions for x to summarize the expected response of measured patience toatastrophe. The respondent selects x to satisfy the marginal condition:

u(c0 + x) + ıE[u(c1)] = u(c0) + ıE[u(c1 + m)] (1)

hich, can be rewritten as

u(c0 + x) − u(c0) = ıp[u(c1 + m) − u(c1)]. (2)

y noting that u(0) = 0 and E[u(c1)] = pu(c1) + (1 − p)u(c0). Eq. (2) summarizes how expectations and curvature should affecthe respondents’ selection of x. If we take as a baseline ı = 1 and p = 1 and linear utility, then choice that satisfies the marginalondition is x = m. If any of the conditions ı < 1, p < 1, or with concave utility c0 < c1 hold, then (x/m) < 1 and our matching taskill code an individual as impatient. If ı < 1, individuals are truly impatient. If p < 1, individuals apply a discount that reflects

heir beliefs about survival. Last, if more consumption will be available in the future (c0 < c1), then, due to concave utility,he benefit from additional future consumption is less than the benefit from additional current consumption. We use thisndifference condition to develop predictions for how the disaster should affect our measure of time preference.

rediction 1. Decreasing survival expectations reduces measured patience: (∂x)/(∂p) > 0.

If the tsunami caused individuals to assign a lower subjective probability to receiving future consumption, then, to satisfyheir indifference condition, individuals should move consumption from the future to the present. Using the implicit functionheorem to differentiate Eq. (2) with respect to p we see that:

∂x = ı[u(c1 + m) − u(c1)]> 0

∂p u′(c0 + x)

If the effect of the tsunami was to decrease the subjective probability of survival for the affected, then our indifferenceondition predicts a decrease in p and, consequently, in measured patience. It is worth noting, however, that it is possiblehat respondents believe there is some arrival rate for catastrophe and having survived the tsunami makes respondents feels though they are much less likely to encounter calamity in the future. However, Cameron and Shah (2013) show that thesunami greatly increased individuals assessments that another tsunami would occur in the future so, if anything, we shouldxpect p to decrease. To provide additional evidence against the relevance of p, we control for this in our regressions using

204 M. Callen / Journal of Economic Behavior & Organization 118 (2015) 199–214

measures of individuals’ subjective expectations about the future.8 Introducing measures of subjective expectations on theright hand side does not affect our result.

An additional confound for measures of patience is difference in marginal utility of consumption over time. To derivepredictions, we let consumption and the survey response x depend on whether or not an individual experienced a disasterso that {ct(1), x(1)} is consumption and the desired payment in period t ∈ {0, 1} in the affected state and {ct(0), x(0)} isconsumption and the desired payment in the unaffected state.

Prediction 2. If pre-event consumption is flat or increasing over time (c0(0) ≤ c1(0)), more consumption is lost in period 0 thanis lost period 1 (c0(1) − c0(0) ≤ c1(1) − c1(0)), and the individual is weakly patient (x(0) ≤ m), then concavity of the utility functionimplies that x should decrease in response to disaster.

This result follows directly from the concavity of the utility function. To see this, note that

[U(c0(1)) + x(0)) − u(c0(1))] − [U(c0(0) + x(0)) − u(c0(0))] ≥ [U(c1(1)) + m) − u(c1(1))] − [U(c1(0) + m) − u(c1(0))]

> ıp[U(c1(1)) + m) − u(c1(1))] − [U(c1(0) + m) − u(c1(0))]

where the first weak inequality follows from our assumptions and the concavity of the utility function and the second holdsif either p or ı is less than 1. Canceling equal terms based on our marginal indifference condition (Eq. (2)) from the first andthe third expression in the inequality, we have that

[U(c0(1)) + x(0)) − u(c0(1))] > ıp[U(c1(1)) + m) − u(c1(1))]

which directly implies x(1) < x(0) to satisfy indifference. Thus we have that x will adjust downward to satisfy our marginalindifference condition in response to a catastrophe. This corresponds to our intuition that, if the effect of the tsunami is tomove respondents with concave utility functions to a lower and weakly steeper consumption path, the resulting increase inthe marginal utility of current consumption should encourage affected respondents to have an increased taste for currentconsumption.

3.2. Predictions of increased patience

Prediction 3. Increases in the pure rate of time preference should increase elicited discount factors: (∂x)/(∂ı) > 0.

This result simply establishes that some of what our elicitation procedure measures is what we are interested in. To showthat the relationship between time discounting and time preference is positive and linear we take the partial derivative withrespect to the rate of pure time preference. To obtain this result, we just need to differentiate Eq. (2) with respect to ı.

∂x

∂ı= p[u(c1 + m) − u(c1)]

u′(c0 + x)> 0.

4. Data and preference measurement

Our data come from a survey of Sri Lankan workers residing in regions affected by the Tsunami undertaken two and ahalf years after the tsunami in July 2007.9 456 wage workers comprise our sample. 155 (34%) of the workers in our surveylost some household assets as a result of the Tsunami. Table 1 reports summary statistics of variables that are affected bythe tsunami and Table 2 provides evidence for balance on endline observables and retrospective variables which should beorthogonal to tsunami exposure if our identifying assumption is correct. The survey of wage workers we use, which elicitedmeasures of patience using a matching task, are described in detail in de Mel et al. (2008). The data were collected two anda half years after the event, which suggests that the changes in preferences we observe are enduring.

Our data suggest that economic recovery from the tsunami was well underway but not complete. The median affectedrespondent indicates having repaired about 50% of the damage resulting from the Tsunami. The data suggest that part of therecovery resulted from the rapid large scale international aid response. Of the 155 individuals who suffered some damage,133 received a relief grant and 140 received some type of recovery aid. On average, aid to the affected workers in our sample

was equal to 87.2% of reported losses. We strongly reject, however, that received aid was equal to the value of the damages(p-value = 0.0002). We additionally check a range of specifications which include both the rupee value of damages and theestimated percentage of assets replaced as controls and find that the effect of disasters on preference remains robust.

8 Our optimism index is the first principal component of subjects’ indication of the degree they agree with the following three statements (measuredon a five point Likert scale where 5 means “strongly agree” and “1 means strongly disagree”): (1) “In uncertain times I usually expect the best”; (2) “I’malways optimistic about my future”; (3) “Overall I expect more good things to happen to me than bad”.

9 Suresh de Mel, David McKenzie, and Christopher Woodruff were the Principal Investigators for the project that yielded these data, which they havekindly shared with us. Their study aimed to explore and document the process of post-tsunami reconstruction (de Mel et al., 2011). The survey andlaboratory tasks were translated by Professor de Mel, who is a native Sri Lankan. An advantage of using their data is that it provides remarkable geographiccoverage and was obtained in the immediate aftermath of the study.

M. Callen / Journal of Economic Behavior & Organization 118 (2015) 199–214 205

Table 1Summary statistics for variables related to tsunami impact.

Variable Full sample Not affected Affected Difference p-Value

Average discount factor 0.819 0.802 0.853 0.052 0.001[0.155] [0.164] [0.131] (0.015)

Distance to coast (km) 17.685 18.145 16.775 −1.371 0.259[12.241] [12.694 [11.275] (1.214)

Elevation (100 m) 1.033 1.038 1.022 −0.016 0.946[2.331] [2.827] [0.622] (0.231)

Share damaged in same grid 0.336 0.172 0.660 0.488 0.000[0.33] [0.223] [0.262] (0.023)

Monthly wage (1000 rupees) 9.970 10.273 9.370 −0.904 0.146[6.266] [6.470] [5.813] (0.621)

Percent of damage repaired 20.349 0.000 60.647 60.647 0.000[32.811] [0.000] [27.616] (1.585)

Damages (1000 rupees) 168.911 0.000 507.851 507.851 0.000[727.179] [0.000] [1193.174] (68.470)

Recovery funds (1000 rupees) 51.203 0.000 152.606 152.606 0.000[135.651] [0.000] [198.762] (11.406)

Coefficient of relative risk aversion 1.154 1.440 0.581 −0.858 0.177[6.343] [6.481] [6.037] (0.636)

Notes: Data are from a survey of wage workers conducted in July 2007. The table reports means, with corresponding standard deviations in brackets andstandard errors in parentheses. Elevations are calculated using the United States Geological Survey Center for Earth Resources and Observation Sciences(EROS) 30 arc second × 30 arc second (approximately 1 km) Digital Elevation Model. The point estimate corresponding to the mean difference betweensubjects affected and not affected by the tsunami for the coefficient of relative risk aversion suggests that the event decreased risk aversion, though this isnot statistically significant. Cameron and Shah (2013) and Cassar et al. (2013) find a statistically significant increase in risk aversion. The data reflect 456subjects in total. 153 of the 456 subjects were affected by the tsunami.

Table 2t-Tests of equality for affected and unaffected workers.

Not damaged Damaged Difference p-Value(ND) (D) (D) − (ND) h0 : (D) = (ND)

In Same Job Since Before Tsunami (=1) 0.683 0.719 0.036 0.434[0.466] [0.451] (0.046)

Gender (Female = 1) 0.370 0.333 −0.036 0.446[0.484] [0.473] (0.048)

Years of education 10.403 10.510 0.107 0.719[3.107] [2.770] (0.297)

Household size 4.495 4.549 0.054 0.741[1.715] [1.504] (0.163)

Marital status (married = 1) 0.690 0.752 0.062 0.170[0.463] [0.433] (0.045)

Age 37.096 38.497 1.401 0.231[11.797] [11.726] (1.168)

Digit span recall 6.602 6.376 −0.226 0.135[1.571] [1.368] (0.151)

Father’s years of education 7.543 7.976 0.433 0.242[3.504] [3.171] (0.369)

Sinhalese (=1) 0.937 0.961 0.023 0.299[0.243] [0.195] (0.023)

English speaker (=1) 0.142 0.124 −0.018 0.603[0.350] [0.331] (0.034)

Tamil speaker (=1) 0.069 0.059 −0.010 0.671[0.254] [0.236] (0.025)

Hindu (=1) 0.007 0.000 −0.007 0.315[0.081] [0.000] (0.007)

Muslim (=1) 0.056 0.033 −0.023 0.271[0.231] [0.178] (0.021)

Buddhist (=1) 0.911 0.922 0.011 0.701[0.285] [0.270] (0.028)

Wage work pre-tsunami 0.604 0.667 0.063 0.192[0.490] [0.473] (0.048)

Casual worker pre-tsunami 0.370 0.314 −0.056 0.238[0.484] [0.466] (0.047)

Self-employed pre-tsunami 0.066 0.052 −0.014 0.565[0.249] [0.223] (0.024)

Apprentice pre-tsunami 0.033 0.033 −0.000 0.985[0.179] [0.178] (0.018)

Worked overseas pre-tsunami 0.023 0.026 0.003 0.842[0.150] [0.160] (0.015)

Notes: Data are from a survey of wage workers conducted in July 2007. Standard deviations are in brackets and standard errors are in parentheses.

206 M. Callen / Journal of Economic Behavior & Organization 118 (2015) 199–214

The workers in our sample are young, predominately male, Sinhalese and Buddhist. They appear to have been impov-erished during their childhood. The average monthly wage in our sample is about 10,000 rupees which, at exchange ratescontemporary to the survey, translates to roughly US $3.30 a day. Our results apply only to this demographic. However, akey focus of the development literature are the reasons individuals transition from wage work to entrepreneurship, andso it may be valuable to understand the responsiveness of time-preference to shocks in this population. Given the rangeof studies that suggest that life experience can permanently alter behavior, we believe these results have some generaliz-ability. Wage workers, moreover, may be constrained with regard to selecting where they live, which allays some concernthat pre-tsunami time preference influences exposure to the Tsunami. We provide a full description of our time preferenceelicitation procedure in Appendix Section A6.

5. Empirical strategy

We take two approaches to identifying the effect of disasters on time preference. First, we provide evidence that exposureto the tsunami is exogenous with respect to pre-event time preference and then estimate the mean difference betweenaffected and unaffected workers, which, if our identifying assumption is valid, can be interpreted as causal. Second, we usea natural discontinuity created by the high water mark, GPS coordinates for our respondents, and GIS data to estimate theeffect using a regression discontinuity approach. Both strategies provide highly significant and comparable estimates, and,in both cases, we observe an increase in measured patience.

The purpose of the framework developed in Section 3 above is to delineate the mechanisms through which experiencinga disaster could affect measures of patience. Prediction 1 indicates that an increase in p – the subject’s belief about theprobability that they will receive the payment in one month – should lead to a decrease in our declarative measure of patience.Prediction 2 indicates that if the effect of the tsunami was to reduce consumption and put subjects on a steeper temporalconsumption profile as they rebuild their assets, then exposure should also lead to an increase in patience. Prediction 3, bycontrast, indicates that if the effect of the tsunami was to increase a subject’s discount factor, ı, then this should also increaseour measure of patience.10

5.1. Approach 1 – Comparing mean differences assuming random assignment to exposure

In this approach, our identifying assumption is that suffering damage from the tsunami is random. This assumption hastwo testable implications. First, if exposure is random, then our sample should be balanced on retrospective variables and onfixed variables which should not be altered by tsunami exposure. If these observables are correlated with time preference,then the absence of any statistical difference between affected and unaffected workers we observe in Table 1 supportsexogeneity. Second, there should be no correlation between preferences and tsunami vulnerability in places that were notaffected by the tsunami because of topography and the direction of the wave, but could have been had the wave had adifferent point of origin.

Our identifying assumption will be violated if individuals select where they live based on their preferences. This mayplausibly happen if preferences influence selection into different lines of work and if the spatial distribution of jobs createsdifferences in exposure probability. Table 2 provides evidence in support of our identifying assumption. We find that oursample is strongly balanced for variables that are plausibly exogenous to exposure or that are retrospective. Age, gender, andeducation, which are among the strongest and most consistent predictors of patience across studies, appear to be balanced.It is also important to note that we find no statistical difference for the type of pre-tsunami employment contract. Laffontand Matoussi (1995) provide evidence that the selection of employment contracts depends on preferences. Moreover, aplausible explanation for selection into tsunami vulnerability according to preferences is through job selection. We interpretthe absence of any statistically significant difference in the type of employment contract prior to the Tsunami as evidencesuggesting that average time preference was identical in both groups prior to the Tsunami.

Figs. 1 and 2 provide more intuition for our first approach to identifying the response of preferences to catastrophe. Fig. 1depicts the location of survey respondents on a topographical map overlaid with 0.07 by 0.07 arc degree fishnet grids.11

The fishnet grids allow us to test whether our result is robust to using only within variation for small regions less subjectto concerns about migration and pre-event selection of location on preference. In the next section, we also make use ofthe fishnet grids to implement a regression discontinuity test of the response of preferences to the tsunami. We see that,while damages were highly concentrated, they do not appear to be consistently correlated with distance to the coast or

elevation owing to the origination of the wave off the coast of Sumatra to the northeast. In the eastern parts of the sample,for example, there are clusters of respondents who live next to the coast at low elevation who were not affected because of thewave’s direction. This is in line with the emphasis placed on bathymetry and topography in the literature that tries to modeland predict tsunami inundation (see e.g. Dominey-Howes and Papathoma, 2007 and Koshimura et al., 2008). Fig. 2 plots

10 Appendix Section A1 also directly tests Predictions 1 and 2. Only Prediction 2 receives support. Our proxy measure for p, the optimism index, is notsignificantly correlated with our time preference measure. Income is positively correlated with our time preference measure, suggesting that a reductionin income would create a reduction in measured preference. We interpret this with great caution as these measures are clearly highly endogenous.

11 Dividing the space where this survey was administered into 0.07 arc degree fishnet grids creates 87 units with an average of 5.52 respondents per grid.

M. Callen / Journal of Economic Behavior & Organization 118 (2015) 199–214 207

Fig. 1. 0.07 by 0.07 arc degree fishnet grids.

Fig. 2. The spatial distribution of discount factors.

208 M. Callen / Journal of Economic Behavior & Organization 118 (2015) 199–214

Table 3Tsunami vulnerability and discounting for unaffected workers.

Dependent variable Average discount factor, not damaged (ND) sample(1) (2) (3) (4) (5)

Distance to coast (km) 0.001 0.001 0.004(0.001) (0.001) (0.006)[0.001] [0.001] [0.005]

Elevation (100 m) −0.000 0.001 −0.026(0.001) (0.002) (0.028)[0.001] [0.002] [0.029]

Share damaged in same grid 0.016 0.020(0.039) (0.039)[0.041] [0.041]

Constant 0.787*** 0.802*** 0.799*** 0.783*** 0.761***

(0.016) (0.009) (0.012) (0.019) (0.100)[0.027] [0.016] [0.017] [0.029] [0.083]

Fishnet grid effects No No No No Yes

R-squared 0.004 0.000 0.000 0.004 0.407# Observations 303 303 303 303 303

p-value (joint significance)No clustering 0.258 0.952 0.676 0.466 0.617GN clustered SEs 0.373 0.954 0.687 0.674 0.645

Notes: This table reports on the relationship between predictors of tsunami exposure and time preferences among wage workers unaffected by the tsunami.All specifications are Ordinary Least Squares. Fishnet grid effects are separate dummy variables for each of the . 07 × .07 fishnet grids overlaying our sample.The share damaged in the same grid is calculated as the share of respondents living in the same grid as the respondent who have been impacted by thetsunami. Data are from a survey of wage workers conducted in July 2007. Robust standard errors reported in parentheses and robust standard errorsclustered at the Grama Niladara (GN) level reported in brackets. There are 52 Grama Niladara in our data. The average discount factor is the average ofresponses to four hypothetical survey questions: the amount required today to forego m rupees in one month and the amount required in 5 months toforego m in 6 months where m ∈ {5000, 10, 000} rupees.Level of significance:

* p < 0.1.** p < 0.05.

*** p < 0.01.

elicited time preferences on a topographical map. In this figure, there does not appear to be an obvious correlation betweenelevation or distance to the coast among unaffected populations, which is consistent with the idea that individuals do notselect vulnerability based on their discount rates. We now turn to a formal test of whether preferences and vulnerability arecorrelated.

In Table 3, we report the results of a regression of our elicited discount factor on the straight-line distance between theindividual and the coast, the elevation of the respondent’s household, and the average number of respondents in the same0.07 by 0.07 arc degree fishnet grid affected by the tsunami. We run this test only for unaffected individuals, as we argue thatthe experience of the tsunami creates a relationship between patience and vulnerability in the affected sample. A failure toreject the null hypothesis that these measures do not describe any differences in elicited patience in our sample is evidencein support of our identifying assumption. To make this test as stringent as possible, we report both standard errors clusteredat the Grama Niladara level and standard errors with no clustering and additionally report the p-values corresponding toan F-test for joint significance for both sets of standard errors.12 In no cases do we find evidence that discount factors andvulnerability are correlated in our unaffected sample.

Given this evidence, there is a case, which we reinforce using regression discontinuity in the next section, that unaffectedworkers provide a valid counterfactual for workers affected by the tsunami.

Thus, we can compare the average time preference of workers affected by the Tsunami with those of workers unaffectedby the Tsunami to determine the causal effect of exposure on time preference. This motivates the specification:

Average Discount Factori = ˇ0 + ˇ1Damagedi + ˇ2Xi + �i (3)

where Average Discount Factori is the discount factor described in the previous section, Damagedi is a dummy variable equalto 1 if the Tsunami destroyed household assets belonging to worker i, and Xi are controls.

5.2. Approach 2 – Regression discontinuity

While the assumptions necessary to identify a causal effect by comparing means are supported by the data, becausewe lack experimental assignment to catastrophe and pre-event baseline data to check balance, we additionally estimatethe effect of the tsunami on time preference using regression discontinuity. This approach ameliorates some additional

12 There are an average of 15.2 respondents pre Grama Niladara and 52 Grama Niladara units in our data

M. Callen / Journal of Economic Behavior & Organization 118 (2015) 199–214 209

Table 4The effect of tsunami exposure on elicited discount factors.

Dependent variable Average discount factor

(1) (2) (3) (4) (5) (6) (7) (8) (9)

Damaged (=1) 0.052** 0.052*** 0.054* 0.072*** 0.073*** 0.063** 0.048** 0.050** 0.022(0.020) (0.019) (0.032) (0.018) (0.018) (0.031) (0.019) (0.019) (0.025)

Coefficient of relative risk aversion 0.002** 0.002 0.001 0.001 −0.001 −0.001(0.001) (0.001) (0.001) (0.001) (0.001) (0.001)

Optimism index −0.012 −0.017* −0.010 −0.015* −0.007 −0.012(0.010) (0.009) (0.009) (0.008) (0.008) (0.007)

Elevation (m) 0.000 0.000 −0.000(0.000) (0.000) (0.000)

Distance to coast (km) 0.001 0.001 0.001(0.001) (0.001) (0.004)

Monthly wage (1000 rupees) 0.003** 0.003** 0.002(0.001) (0.001) (0.001)

Percent of damage repaired 0.000 0.000 0.000(0.000) (0.000) (0.000)

Damages (1000 rupees) 0.000** 0.000** 0.000(0.000) (0.000) (0.000)

Recovery funds (1000 rupees) −0.000 −0.000 −0.000(0.000) (0.000) (0.000)

Constant 0.802*** 0.799*** 0.746*** 0.795*** 0.793*** 0.749*** 0.803*** 0.803*** 0.773***

(0.016) (0.016) (0.030) (0.013) (0.013) (0.027) (0.009) (0.009) (0.069)Fixed Effects No No No DS DS DS FN FN FNR-squared 0.025 0.042 0.076 0.073 0.088 0.116 0.345 0.360 0.377# Observations 456 448 446 456 448 446 456 448 446# Clusters 52 52 52 52 52 52 52 52 52

Notes: This table reports on the effects of exposure to the tsunami on our average discount factor measure. Data are from a survey of wage workers conductedin July 2007. All specifications are Ordinary Least Squares. Standard errors clustered at the Grama Niladara level are reported in parentheses. There are 52Grama Niladara in our data. We calculate the optimism index as the first principal component of three questions related to expectations about the future.The fixed effect samples are DS = District (7 divisions) and FN = 0.07 × 0.07 arc degree Fishnet Grids (87 divisions). The coefficient of relative risk aversion iselicited using incentivized Holt–Laury price lists. The average discount factor is the average of responses to four hypothetical survey questions: the amountrequired today to forego m rupees in one month and the amount required in 5 months to forego m in 6 months where m ∈ {5000, 10, 000} rupees.Level of significance:

tsi

utmba

wad

6

6

da

* p < 0.1.** p < 0.05.

*** p < 0.01.

hreats to our identification, such as whether differential exposure to aid workers affected how individuals respond to aurvey. Assuming that individuals were not able to perfectly anticipate the high water mark, the individuals who live eithermmediately above or below the water mark should be balanced on unobservables.

Regression discontinuity estimates, therefore, are unlikely to represent pre-existing differences in preferences or in othernobservables. To obtain the estimate, we first calculate the elevation for each of the respondents in the sample based onheir GPS coordinates. Because data on the precise high watermark in all locations do not exist, we identify the high-water

ark as the highest elevation at which a respondent reports being affected both in each Grama Niladara and in each 0.07y 0.07 arc degree fishnet grid. We show below that our approach is robust to using variation both within pre-existingdministrative units (Grama Niladara) and evenly sized arbitrary fishnet grids.

We estimate the effect of tsunami exposure on time preference using the following regression:

Average Discount Factori = �0 + �1Damagedi + �2 Mi + �3M2i + �4M3

i + �5M4i + �6Xi + �i (4)

here Mi is the distance in meters between individual i and the local high water mark and Damagedi is instrumented with dummy variable equal to one when respondent i is below the recorded water level (Mi < 0) to provide a fuzzy regressioniscontinuity estimate.13

. Results

.1. Approach 1 – Estimates of mean differences under random assignment to exposure

Table 4 reports results for Specification 3. We cluster standard errors at the Grama Niladara level to account for the highegree of spatial correlation in tsunami exposure. Our estimates are consistent with the tsunami creating an increase in theverage monthly discount factor from about 0.8 to about 0.85 or to about 0.9 depending on the specification. This represents

13 In results not reported here, we show that our results are not changed using a sharp discontinuity design.

210 M. Callen / Journal of Economic Behavior & Organization 118 (2015) 199–214

Fig. 3. Regression discontinuity evidence. Note: 4th-order polynomial regressions with 95% confidence intervals reported. High water marks for each ofthe 52 Grama Niladara administrative units in the sample are calculated as the elevation of the highest individual in the sample reporting damage fromthe tsunami. We exclude from this figure the 52 respondents residing exactly at the high water mark.

roughly a 1/3rd to 2/3rd standard deviation increase in our patience measure. While problems with taking the cardinality ofthis measure seriously are well-known (see e.g. Andreoni and Sprenger, 2012; Frederick et al., 2002), this provides evidenceof patience increasing as a result of the event.

We attempt to control for locational sorting first by including elevation and distance to the coast as proxies for vul-nerability in the regression and by using two different sets of fixed effects. In some specifications we add our measure ofrisk preference elicited using incentivized Holt–Laury MPLS as these are argued to describe some of the variation in timepreference (Harrison et al., 2008).14 Columns 1–3 report results with no fixed effects, columns 4–6 provide results whendummies are added for the 7 districts represented in the data and columns 7–9 include dummies for 87 arbitrary evenlysized 0.07 arc degree fishnet grids. We find that our result is robust except for in column 9, which is quite demanding on thedata as we are attempting to estimate 9 parameters using within variation for geographic divisions that have an average of5.2 respondents.

6.2. Approach 2 – Regression discontinuity estimates

Because we lack pre-event data and random assignment, we use data on respondent locations and an estimate of the highwater mark to estimate the effect of the tsunami on time preference using regression discontinuity. Fig. 3, which depicts thedistribution of respondents’ elevation difference from the high watermark excluding the one worker per GN who resides atexactly the high water mark. Consistent with the tsunami being an imprecisely predicted disaster, we do not observe anyclear bunching near the watermark. Fig. 4 graphs the average elicited discount factor and the estimated 4th-order polynomialalong with 95% confidence intervals when the high water mark is calculated within Grama Niladara unit. The continuousdistribution in Fig. 3 and the visually salient discontinuity in Fig. 4 suggest that we can estimate the causal effect of thetsunami on time preference using a regression discontinuity design.

Table 5 reports results for specification 4. Columns 1 and 2 provide estimates based on calculating the high water markas the highest elevation within a Grama Niladara division in which a respondent reports damage. Hydrological modelsof inundation and using satellite imagery to estimate the extent of damage are known to be highly imperfect and highlysensitive to assumptions about wave sizes at landfall (see e.g. Dominey-Howes and Papathoma, 2007; Koshimura et al.,2008). To check the robustness of our result to additional estimations of the high water mark, columns 3 and 4 report resultswhere the high water mark is calculated as the elevation of the highest respondent reporting tsunami damage within small

15

fishnet grids (0.05 by 0.05 arc degrees) of about 5 km by 5 km. Even in column 3, where the result is not significant atconventional levels, we obtain a p-value of 0.114.16 Fig. 5 depicts the histogram of time preference measures for affectedand unaffected samples. Visually, it appears that the affected sample is more patient.

14 In contrast to the positive correlation between risk and time preference in Column 2, Willinger et al. (2013) find a negative correlation. This correlation,in our data, however, is not very robust. Indeed it is insignificant in all other specifications in Table 5.

15 In results available on request, we find that our specification is robust to different fishnet grid sizes16 In results available on request, we check the robustness of our estimates to taking a multilevel modeling approach to estimation. Taking this approach,

our estimate of the coefficient on the “Damaged (=1)” regressor for specifications (1) to (3) are 0.052 (SE = 0.015), 0.052 (SE = 0.015), and 0.055 (SE = 0.032),respectively.

M. Callen / Journal of Economic Behavior & Organization 118 (2015) 199–214 211

Ftt

6

wttdd

TF

NopihwtwL

*

ig. 4. Regression discontinuity evidence. Note: 4th-order polynomial regressions with 95% confidence intervals reported. High water marks for each ofhe 52 Grama Niladara administrative units in the sample are calculated as the elevation of the highest individual in the sample reporting damage fromhe tsunami.

.3. Interpretation

The increase in patience estimated using a comparison of means ranges from 0.048 to 0.072 (ignoring column (9) wheree lack the power to obtain precise estimates) and the increase estimated using regression discontinuity ranges from 0.077

o 0.108. The similarity of the coefficients obtained using two different approaches, as well as the evidence supportinghe identifying assumption for the first approach and the robustness of those estimates to using within variation from 87ifferent geographic units provides evidence that exposure to the tsunami increased patience. We now review and test theegree to which selective migration might be driving our results.

able 5uzzy regression discontinuity results.

Dependent variable Average discount factor

(1) (2) (3) (4) (5) (6)

Damaged (=1) 0.089* 0.105* 0.085 0.080 0.110*** 0.112***

(0.053) (0.058) (0.062) (0.060) (0.038) (0.042)Distance to water mark (km) −0.000 −0.000 0.000 −0.000 0.000 0.000

(0.000) (0.000) (0.000) (0.000) (0.000) (0.000)Distance to water mark2 −0.000 −0.000 −0.000 −0.000 −0.000 −0.000

(0.000) (0.000) (0.000) (0.000) (0.000) (0.000)Distance to water mark3 0.000 0.000 −0.000 −0.000 −0.000 −0.000

(0.000) (0.000) (0.000) (0.000) (0.000) (0.000)Distance to water mark4 0.000 0.000 0.000 0.000 0.000 0.000

(0.000) (0.000) (0.000) (0.000) (0.000) (0.000)Constant 0.813*** 0.774*** 0.656*** 0.623*** 1.002*** 0.985***

(0.020) (0.027) (0.023) (0.041) (0.005) (0.027)R-squared 0.061 0.097 0.213 0.257 0.325 0.358N 442.000 432.000 442.000 432.000 442.000 432.000Fixed effects No No GN GN FN FNExtra controls No Full No Full No Full

First stage F-statistic 177.54 98.297 107.774 91.470 94.370 53.773R-squared 0.061 0.097 0.213 0.257 0.325 0.358N 442 432 442 432 442 432

otes: This table provides Regression Discontinuity estimates of the impact of the tsunami on time preferences. All specifications include the full setf interacted between affected and the Distance to Water Mark polynomial terms. Standard errors clustered at the Grama Niladara level reported inarentheses. There are 52 Grama Niladara in our data. Damaged is instrumented with a dummy equal to 1 if a respondent lived below high water mark

n their Grama Niladara, which provides the fuzzy regression discontinuity estimate. The average discount factor is the average of responses to fourypothetical survey questions: the amount required today to forego m rupees in one month and the amount required in 5 months to forego m in 6 monthshere m ∈ {5000, 10, 000} rupees. The high water mark is calculated as the highest elevation at which someone within the cluster reports being hit where

he clusters are GN = Grama Niladara and FN = 0.07 × 0.07 arc degree Fishnet Grids. The full set of controls are elevation, distance to the coast, monthlyage, CRRA, optimism index, gender, marriage status, years of education, household size, and age.

evel of significance:* p < 0.1.

* p < 0.05.*** p < 0.01.

212 M. Callen / Journal of Economic Behavior & Organization 118 (2015) 199–214

Fig. 5. Tsunami exposure and the distribution of preferences. Notes: Data are from a survey of wage workers conducted in July 2007. The average discountfactor is the average of responses to four hypothetical survey questions: the amount required today to forego m rupees in one month and the amount

required in 5 months to forego m in 6 months where m ∈ {5000, 10, 000} rupees.

6.4. Selective migration

The survey was conducted 2 and a half years after the event. An alternative explanation therefore is that the impatientindividuals may have migrated out in response to the event, creating a difference in means between affected and unaffectedsamples that does not reflect a causal change. While we cannot rule this out conclusively, several findings suggest that itmay not be a major issue. We address this potential concern using detailed data on the job histories of wage workers in oursample. These histories cover periods prior to the tsunami. With these data, we can check whether the tsunami appears tohave affected a proxy measure of migration, whether a respondent has been in the same job since prior to the tsunami, andthe distribution of workers across jobs.

In Table 2, we see that the share of workers in the same job since before the tsunami is the same for both affectedand unaffected populations. Table 3A reports regressions of a dummy variable equal to one on an indicated for whethera respondent experienced the tsunami. In columns (1) to (3), we find no significant differences. In column (4), we find,using variation within arbitrary 0.07 × 0.07 fishnet grids, that respondents who are damaged are 15 percentage pointsmore likely to have been in the same job since before the tsunami. We obtain a similar, though insignificant, result incolumn (6) using a fuzzy regression discontinuity approach. This indicates that, right at the high water mark, it appearsthat unaffected individuals were more likely to switch jobs. An alternative interpretation is that individuals in the samejob since prior to the tsunami are more likely to have actually experienced it than others living in their immediatevicinity.

To provide an additional, possibly more convincing, check, we create a simple sectoral employment Herfindahl indexdefined as

∑js

2j

where sj is the share of workers employed in a given category j at the time of the survey. We have dataon the following categories: wage work, casual work, self-employed, and apprenticeship. If the tsunami created large-scalemigration either out (because of people leaving decimated areas) or in (as wage workers came to seek new employmentopportunities), then we would expect that the distribution of workers across these positions would shift due to economicchanges created by the tsunami. Running a simple regression of the sectoral employment Herfindahl on whether a dummyvariable equal to one for affected Grama Niladaras suggests no relationship.17 In results available from the author, I also do

not find any difference in average employment within each category.

17 The Herfindahl in unaffected Grama Niladaras is 0.797 (se = 0.087) and in affected Grama Niladaras is 0.806 (se = 0.051). The difference between thesetwo means is −0.009 (se = 0.101) with a corresponding p-value of 0.930. In results available on request, I find no relationship using a battery of regressions,including fuzzy RD specifications run at the respondent level.

6

itoro

7

rrp

aaps

cmfub

A

2

R

AAAA

AAB

B

BBBBBC

CCCCddD

E

F

FFHJ

M. Callen / Journal of Economic Behavior & Organization 118 (2015) 199–214 213

.5. Alternative explanations

In order to establish that our result reflects an increase in pure time preference, we need to show that the effect we finds not the result of some other change in the economy due to the tsunami. The theory we develop above allows us to narrowhe list somewhat. In Appendix Section A3, we consider four alternative explanations and provide evidence consistent withur interpretation. In that section, we address the following potential confounds: (i) the selective migration of exposedespondents; (ii) increases in the demand for savings from exposure; and (iii) problems arising from the hypothetical naturef the measurement protocol.

. Conclusion

Explanations for the discounting of future utility have a long history in economic research. However, only recently haveesearchers had the data to test for systematic causes of heterogeneity in economic preference. This approach, rather thanelying on introspection, forces the researcher to provide credible documentation that the hypothesized determinants ofreference create differences causally.

In this paper we provide evidence that, contrary to a standard modeling assumption, preferences are systematicallyffected by life experience. Our data provide measures of time preference for a population that was severely affected by

major disaster and also permit us to exploit exogenous variation in exposure to that disaster to estimate its’ effect onreferences. These data therefore allow us to attempt a stringent test of whether preference parameters can be linked in aystematic way to life experience.

We find that exposure increases patience using two different empirical strategies. A battery of tests supports that thesehanges are indeed attributable to a change in preferences and not to other changes in the economic environment. Comple-enting this, we find preliminary evidence that experiencing the event appears to substitute for other inputs to preference

ormation such as education. It is our hope that these results continue to open the door for a broader agenda which seeks tonderstand sources of heterogeneity in economic preference and that relies not only on differences in prices and budgets,ut also on systematic and well-understood differences in preferences to explain differences in economic behavior.

ppendix A. Supplementary data

Supplementary data associated with this article can be found, in the online version, at http://dx.doi.org/10.1016/j.jebo.015.02.019.

eferences

lesina, A., Giuliano, P., 2010. Preferences for redistribution. Handbook of Social Economics, vol. 1A. North Holland, pp. 93–131.ndrabi, T., Das, J., 2010. In Aid We Trust: Hearts and Minds and the Pakistan Earthquake of 2005. Working Paper.ndreoni, J., Sprenger, C., 2012. Estimating time preferences from convex budgets. Am. Econ. Rev. 102 (7), 3333–3356.shraf, N., Karlan, D., Yin, W., 2006. Tying Odysseus to the mast: evidence from a commitment savings product in the Philippines. Q. J. Econ. 121 (May (2)),

635–672.ubert, R., Aubert, C., 2013. Do Natural Disasters Modify Risk Preferences? Evidence from a Field Experiment in Vietnam. Working Paper.ugenblick, N., Niederle, M., Sprenger, C., 2013. Working Over Time: Dynamic Inconsistency in Real Effort Tasks.auer, M., Chytilova, J., Morduch, J., 2012. Behavioral foundations of microcredit: experimental and survey evidence from rural India. Am. Econ. Rev. 102

(November (2)), 1118–1139.chir, M.A., Willinger, M., 2013. Does the exposure to natural hazards affect risk and time preferences? Some insights from a field experiment in Peru.

Working Paper.ecker, G.S., Mulligan, C., 1997. The endogenous determination of time preference. Q. J. Econ. 112 (3), 729–758.en-Ner, A., Putterman, L., 2001. Trusting and trustworthiness. Boston Univ. Law Rev. 81, 523–551.lanchard, O.J., 1985. Debt, deficits, and finite horizons. J. Polit. Econ. 93 (2), 223–247.lattman, C., Annan, J., 2010. The consequences of child soldiering. Rev. Econ. Stat. 92 (4), 882–898.öhm-Bawerk, E.V., 1889. Capital and Interest. Libertarian Press, South Holland.allen, M., Isaqzadeh, M., Long, J.D., Sprenger, C., 2014. Violence and risk preference: experimental evidence from Afghanistan. Am. Econ. Rev. 104 (1),

123–148.ameron, L., Shah, M., 2013. Do Natural Distasters Shape Risk Attitudes? Working Paper.assar, A., Healy, A., von Kessler, C., 2013. Trust Risk and Time Preferences After a Natural Disaster: Experimental Evidence from Thailand. Working Paper.astillo, M., Carter, M., 2011. Behavioral Responses to Natural Disasters. Working Paper.huang, Y., Schechter, L., 2014. Stability of Experimental and Survey Measures of Risk, Time, and Social Preferences Over Multiple Years. Working Paper.e Mel, S., McKenzie, D., Woodruff, C., 2008, February. Who are the Microenterprise Owners? Evidence from Sri Lanka on Tokman v. de Soto.e Mel, S., McKenzie, D., Woodru, C., 2011. Enterprise recovery following natural disasters. Econ. J. 122, 64–91.ominey-Howes, D., Papathoma, M., 2007. Validating a Tsunami Vulnerability Assessment Model (the PTVA Model) using field data from the 2004 Indian

Ocean Tsunami. Nat. Hazards 40, 113–136.ckel, C.C., El-Gamal, M.A., Wilson, R.K., 2009. Risk loving after the storm: a Bayesian-Network study of Hurricane Katrina evacuees. J. Econ. Behav. Organ.

69 (February (2)), 110–124.inlay, J.E., 2009, March. Fertility response to natural disasters: the case of three high mortality earthquakes. Policy Research Working Paper Series 4883,

The World Bank.

isman, R., Jakiela, P., Kariv, S., 2013. How Did Distributional Preferences Change During the Great Recession.rederick, S., Loewenstein, G., O’Donoghue, T., 2002. Time discounting and time preference: a critical review. J. Econ. Lit. 40 (June (2)), 351–401.arrison, G.W., et al., 2008. Eliciting risk and time preferences. Econometrica 76 (3), 583–618.

ayachandran, S., Lleras-Muney, A., 2009. Life expectancy and human capital investments: evidence from maternal mortality declines. Q. J. Econ. 124(February (1)), 349–397.

214 M. Callen / Journal of Economic Behavior & Organization 118 (2015) 199–214

Jevons, H.S., 1905. Essays on Economics. Macmillan, London.Jevons, W.S., 1888. The Theory of Political Economy. Macmillan, London.Karlan, D., 2005. Using experimental economics to measure social capital and predict real financial decisions. Am. Econ. Rev. 95 (5), 1688–1699.Koshimura, S., Oie, T., Yanagisawa, H., Imamura, F., 2008. Vulnerability Estimation in Banda Aceh Using the Tsunami Numerical Model and the Post-Tsunami

Survey Data.Laffont, J.J., Matoussi, M.S., 1995. Moral Hazard, Financial Constraints and Sharecropping in El Oulja. Rev. Econ. Stud. 62 (3), 381–399.Malmendier, U.M., Nagel, S., 2011. Depression babies: do macroeconomic experiences affect risk taking? Q. J. Econ. 126 (1), 373–416.Meier, S., Sprenger, C., 2015. Temporal stability of time preferences. Rev. Econ. Stat. 97 (2), 273–286.Mischel, W., Shoda, Y., Rodriguez, M., 1989. Delay of gratification in children. Science 244 (4907), 933–938.Olson, M., Bailey, M.J., 1981. Positive time preference. J. Polit. Econ. 89 (1), 1–25.Pörtner, C., 2008. Gone With the Wind? Hurricane Risk, Fertility and Education. Working Papers UWEC-2006-19-R, University of Washington, Department

of Economics.Rae, J., 1834. The Sociological Theory of Capital. Macmillan, London.Senior, N.W., 1836. An Outline of the Science of Political Economy. Clowes and Sons, London.Stigler, G.J., Becker, G.S., 1977. De Gustibus Non Est Disputandum. Am. Econ. Rev. 67 (March (2)), 76–90.

Voors, M., Nillesen, E., Verwimp, P., Bulte, E., Lensink, R., van Soest, D., 2012. Does conflict affect preferences? Results from field experiments in Burundi.

Am. Econ. Rev. 102 (2), 941–964.Willinger, M., Bchir, M.A., Heitz, C., 2013. Risk and Time Preferences Under the Threat of Background Risk: A Case Study of Lahars Risk in Central Java.

Working Paper.Yaari, M.E., 1965. Uncertain lifetime, life insurance, and the theory of the consumer. Rev. Econ. Stud. 32 (2), 137–150.